FedMVP: Federated Multi-modal Visual Prompt Tuning for Vision-Language Models [ICCV 2025]

Mainak Singha, Subhankar Roy, Sarthak Mehrotra, Ankit Jha, Moloud Abdar, Biplab Banerjee, Elisa Ricci

Official implementation of the paper "FedMVP: Federated Multi-modal Visual Prompt Tuning for Vision-Language Models"

How to install

Create your environment:

$ conda create -n fedmvp python=3.8
$ conda activate fedmvp
$ conda install pytorch==1.12.1 torchvision==0.13.1 cudatoolkit=10.2 -c pytorch
$ pip install -r requirements.txt

Data preparation:

Please refer to CoOP for data preparation.

Training and Evaluation

Please run the command for training the model:

python Launch_FL.py --root YOUR_DATA_PATH --exp_name \$1 --model_name \$2

--exp_name specifies the generalization setting e.g. cross_cls = base-to-new generalization, multisource_singletarget_office = MSST task in the domains of OfficeHome dataset. Please refer to the config/utils.py file for more details. --model_name refers to the training model e.g. fedmvp

Evaluating the trained models

1. To evaluate the trained FedMVP model for the experiment cross_cls:

python Launch_FL.py --root YOUR_DATA_PATH --exp_name \$1 --model_name \$2 --eval-only --model-dir output/cross_cls/fedmvp/20_8/42/ --load-epoch 200

2. To evaluate the trained FedMVP model for the experiment cross_data:

python Launch_FL.py --root YOUR_DATA_PATH --exp_name \$1 --model_name \$2 --eval-only --model-dir output/cross_data/fedmvp/20_8/42/ --load-epoch 200

Citation

If you use our work, please consider citing:

@article{singha2025fedmvp,
  title={FedMVP: Federated Multi-modal Visual Prompt Tuning for Vision-Language Models},
  author={Singha, Mainak and Roy, Subhankar and Mehrotra, Sarthak and Jha, Ankit and Abdar, Moloud and Banerjee, Biplab and Ricci, Elisa},
  journal={arXiv preprint arXiv:2504.20860},
  year={2025}
}

Acknowledgements

Our implementation builds upon the CoOp, FedTPG and classify_by_description repositories, and we sincerely thank the authors for making their code publicly available.